Hacker News Reader

Comment by duskwuff | Hacker News Reader

Is it actually scanning for signs, or is it using map data to determine the speed for the current road segment? I suspect it's actually the latter; reading road signs is a somewhat difficult CV problem, whereas having a well-annotated map is largely a solved problem already. (My old Garmin GPS had this data!)

View on HN

They very well may be doing the latter as well, but Tesla has specifically said that their camera reads the speed from road signs[1]

[1] http://www.teslamotors.com/blog/dual-motor-model-s-and-autop...

Reading road signs is exceedingly tricky. It's not just the OCR, there's plenty of other gotchas. e.g. what if you miss a road sign because you were overtaking a truck when the sign went past?

Road signs also have situational contexts. e.g. in the UK there are road signs indicating a max speed for an upcoming sharp bend (or series of bends) in the road. There isn't always a matching speed limit sign after the bend because drivers are assumed to work that out for themselves. But will the cars do so? Or will your self-driving car get stuck at the lower speed?

Speed limits almost always have repeater signs at regular intervals, particularly when it's not national speed limit. The recommended max speed signs on bends are different from speed limit signs too. If there's not a red circle, it's not mandatory. The exception is national speed limit signs, which would require the car to see if it's a dual carriageway or not so it can switch between 60 and 70.

In my jurisdiction the yellow "limit" signs before corners are not actually speed limits, just references. You are allowed to disobey them at your peril. Sometimes that means you can do 80kph around a "30" corner, a common situation on mountain roads.

That also avoids the enforcement problems such as where the new limit should being and end, and the difficulty of measuring vehicle speed through a corner.

> Reading road signs is exceedingly tricky.

You (and parents) vastly overestimate difficulty. They are reflective, high contrast, using known font and very limited symbols (0-9). Car CV systems can distinguish pedestrians from background, they can read a freaking sign.

You have to distinguish the sign from signs on trucks (indicating their maximum speed), reflections of signs in windows, you have to recognize the sub-signs which might limit the scope of the main sign (speed limit might only apply to exit or under certain weather conditions), then there are the usual adverse conditions like rain, reflections, dirt, plants, shadows etc.. Just search for "german traffic sign recognition benchmark" to see what the state of the art is (ConvNet).

It knows where cars and the lanes are. So no it doesn't think there are speed signs in middle of road or on a truck.

Sub signs I'll give you. Don't have any where I live though.

Compared to all the other things it must do reading signs is on the easy list.

So can you cause a pile up by putting up a fake road sign with a speed limit of 0 or 1mph and waiting for a Tesla to come along??

It is actually scanning them with the camera.

The MobileEye system in the Tesla (and other manufacturers) has some of the most sophisticated CV software on the road - not just scanning for road signs, but identifying road markings, lane markings, curbs, obstructions, traffic signals, etc.

That's a good question. He is under the impression that it's actually scanning them, but I honestly don't know (I am aware that the mapping data would be much easier to accomplish).

The map is only the start. No map can accommodate all the local changes and rules (small/temp construction zones) nor will local authorities always forward minor changes (school zones).

Actually, most map data _doesn't_ have speed limit data associated with it: http://opendata.stackexchange.com/questions/675/data-source-...

Tesla is most likely using their machine vision system to identify speed limit signs, and then incorporate this data into their mapping data.

Reply on news.ycombinator.com